Abstract: With the rapid development of 3D vision, large-scale 3D point cloud processing in real time based on deep learning has become a research hotspot. Taking a large-scale 3D point cloud with disordered spatial distribution as the background, this study comprehensively analyzes, introduces and compares the latest progress of deep learning in real-time processing of 3D vision problems. Then, it analyzes in detail and compares the advantages and disadvantages of algorithms in terms of point cloud segmentation, shape classification and target detection. Further, it briefly introduces the common data sets of point clouds and compares the algorithm performance of different data sets. Finally, the study points out the future research direction of 3D point cloud processing based on deep learning.
Abstract: Blockchain technology has brought new changes to cryptocurrencies and has been widely used. However, it still faces the needs and goals of high throughput, low transaction latency, security, and decentralization. In addition, the willingness of consumption nodes (i.e., transaction providers) is difficult to be mapped into leaders, and block miners are keen on mining competitions, which also leads to intensified centralization and energy consumption. To this end, a new consensus algorithm, PoM (proof-of-market), and its first implementation case, the Achain protocol, are proposed. The algorithm is different from the traditional PoW (proof-of-work) consensus, and its design enables consumer nodes to perform PoW and vote for leader nodes, which not only discretizes the mining, improves decentralization, and reduces energy consumption but also reflects the willingness of consumer nodes. In other words, only the node mainly supported can become the leader. In terms of performance, Achain also improves scalability compared with PoW-type blockchains, and it provides a solution for node storage and optimization, which is called FastAchain. In terms of security, Achain is supplemented by a set of incentive-compatible reward and punishment mechanisms to make malicious nodes have negative revenue expectations, which protects the interests of honest nodes, and Achain can tolerate up to 1/3 of the total network computing power being controlled by the malicious nodes. In order to verify Achain’s performance, a prototype of Achain under a large-scale network is implemented for evaluation. The results show that Achain has achieved expectations, outperformed some mainstream representative blockchain protocols, and maintained positive chain convergence and decentralization.
Abstract: Accurate prediction of the future trend of the commercial sales volume is of great importance to the development and operation of enterprises and the macro-control by the government. Traditional data prediction methods are time-consuming and subjective, while the existing data-driven future business prediction methods do not take into account the diversity of features in the data sets. The data of the commercial sales volume is time-series, which contains a wealth of time window features, lagging historical features, and price change trend features. Previous studies tend to focus only on some of these features, and the integration and enhancement of these features are seldom explored. The prediction accuracy of the existing future business prediction methods still needs to be improved. Therefore, this study proposes a future business forecast method based on multimodal feature aggregation, which firstly preprocesses the commercial sales volume data and then extracts five different groups of time window features and other features of the data set on the basis of feature engineering. In machine learning, the hard voting mechanism is used to select the appropriate model for the training of the five groups of time window features. At the same time, the neural network optimization model is applied to extract the time-series features and forecast results, and then, the dependency relationships between the data set of the sales volume and some features are analyzed. Finally, with the soft voting model, a high-precision forecast of the commercial sales volume is achieved by complete model integration. The experimental results reveal that the proposed method has high prediction accuracy and efficiency, which is greatly better than the existing prediction methods.
Abstract: This study is conducted to measure the level of industrial green development and judge the differences in industrial green innovation ability among provinces. The study constructs a multi-index evaluation system integrating comprehensive attribute values for industrial green development from the perspective of green input factors and green output benefits and proposes an interval multi-attribute measurement method based on the hybrid multi-dimensional cloud model. This method creatively uses the mutual transformation of the interval weight and cloud weight to solve the problem of different weights of multiple indexes. After that, the parent cloud closeness is employed to calculate the industrial green development level, and the cloud projection is applied to measure industrial green innovation ability. Finally, industrial panel data between provinces are used for verification. The results indicate that compared with the results of the general multi-index comprehensive evaluation method, the empirical results of this method are more consistent with the actual situation. This means that the method can not only evaluate and analyze the overall situation of the industrial green development level but also accurately calculate the contribution of each index to judge whether a province or region has the industrial green innovation ability. Therefore, this study can provide substantive suggestions and theoretical decision-making basis for regions to adjust the measurement indexes for the industrial green development level and formulate industrial green development plans.
Abstract: The existing research on battery state of charge (SOC) prediction based on neural networks mostly focuses on the optimization of model structure and related parameters, ignoring the important role of training data. A battery SOC prediction method based on feature selection and data augmentation is proposed to overcome this problem. Specifically, feature engineering is carried out according to the original battery charge and discharge data, and seven features that are most helpful to model prediction are selected by the permutation importance (PI) method; then, Gaussian noise is added to expand the total number of training data samples and thereby achieve the purpose of data augmentation. In the experiment, a bidirectional long short-term memory (Bi-LSTM) network is used as the prediction model, and the Panasonic 18650PF dataset is adopted as the training data. When the standard Bi-LSTM model is employed for prediction, the mean absolute error (MAE) and the maximum error (MaxE) are 0.65% and 3.92% respectively. After feature selection and data augmentation, the MAE and MaxE of model prediction are 0.47% and 2.62% respectively, indicating that the accuracy of the battery SOC prediction model can be further improved by PI feature engineering and the Gaussian data augmentation method.
Abstract: Facial expression recognition is easy to lose a lot of useful feature information during feature extraction and cannot extract more comprehensive facial expression features. In view of these problems, a multi-scale feature fusion network model (DS-EfficientNet) is proposed. The model includes a deep network and a shallow network. The shallow network is used to extract the detailed texture information of facial expressions, and the deep network is used to extract the global information of expressions. An attention mechanism is added to the shallow network to enhance the ability to extract shallow detail information. Finally, feature fusion is performed on channels, and the network can extract more abundant facial expression information after the fusion. In order to reduce the model parameters and improve the generalization performance of the model, the fully connected layer is replaced by a global average pooling layer, and batch normalization is added. The method proposed in this study is tested on Fer2013 and CK+, and the recognition accuracy reaches 73.47% and 98.84%. Experiments show that this method can extract more abundant facial expression information, and the model has a strong generalization ability.
Abstract: Longitudinal tear detection of conveyor belts is one of the important issues in coal mine safety production. In the longitudinal tear detection of mining conveyor belts, insufficient detection accuracy, false detections, and missing detections occur due to insufficient data, diversified damage patterns, and extreme aspect ratios. In this study, an improved YOLOv4 longitudinal tear detection algorithm for conveyor belts is proposed. First, the existing data is expanded by data enhancement to construct a longitudinal tear data set for conveyor belts. Secondly, the variable convolution is added to the backbone network to enhance the feature extraction ability of the model for diverse damage patterns. Finally, in the feature fusion stage, the cross-stage partial network (CSPNet) structure is introduced to improve the longitudinal tear detection performance of the model for extreme aspect ratios, and further reduce missing detection and false detection. The experimental results show that the accuracy of the longitudinal tear detection for the conveyor belt reaches 92.5%, and the F1 score reaches 93.1%, which basically meets the requirements of the longitudinal tear detection for the conveyor belt.
Abstract: To address the poor robustness, low accuracy, and slow operation of existing color-ring resistor recognition methods, this study proposes a lightweight image recognition algorithm for color-ring resistors based on the MobileNetV3 network. Firstly, data augmentation is conducted on a self-built data set to increase the sample size and improve the robustness of the model. Secondly, a convolutional?block?attention?module (CBAM) attention module is utilized in the bottleneck structure, which can enhance the ability of the model to refine features in space and channels for accuracy improvement. Thirdly, the classifier is optimized by removing the redundant dimension-increasing operations, which can reduce the number of parameters while improving accuracy and thereby speed up the operation of the model. Finally, a skip connection is embedded in the network in response to unequal feature image sizes and channel numbers. This makes the model able to extract more features from deep networks and improves accuracy. The experimental results show that the model can recognize color-ring resistors quickly and accurately, with its accuracy reaching 98% on the self-built data set. The model can provide a new technical reference for the automatic recognition of resistors.
Abstract: Existing deep learning-based hashing methods for image retrieval usually cascade several fully connected layers as the hash coding layer and output each bit of the hash code in parallel. This approach treats hash encoding as the information encoding of images and ignores the relevance between bits of the hash code in the coding process and the redundancy of coding, which leads to the limited encoding performance of networks. In light of the principle of code check, this study proposes SHNet, a deep hashing method based on serial encoding. Different from the traditional hashing method, SHNet designs the hash coding network layer structure as a serial mode and verifies the first part of the serial hash codes in the process of generating hash codes, so as to make full use of the relevance and redundancy of codes to generate more informative, more compact, and more discriminative hash codes. Using mAP as the evaluation standard of retrieval performance, the study compares the proposed method with current mainstream hashing methods. The results show that the mAP values of the proposed method under different hash coding lengths are superior to those of the current mainstream deep hashing algorithm on the three datasets of CIFAR-10, ImageNet, and NUS-wide, which proves its effectiveness.
Abstract: Fabric defect detection is an important link in the quality management of the textile industry. Accurate and fast fabric defect detection on embedded devices can effectively reduce the detection cost, thus being of great value. Considering the structural characteristics of colored fabric defects in actual production, such as a complex background, large differences in the quantity of defects, an extremely high aspect ratio, and a high proportion of small defects, a colored fabric defect detection method based on a lightweight model is proposed and implemented on an embedded circuit board Raspberry Pi 4B. The lightweight feature extraction network ShuffleNetV2 is first used to extract the features of colored fabric defects on the basis of the one-stage target detection network, you only look once (YOLO), so as to reduce the complexity of the network structure and the number of parameters and improve the detection speed. Then, the detection head is decoupled to separate the classification and localization tasks so that the convergence speed of the model can be improved. In addition, the complete intersection over union (CIoU) is introduced as the loss function of defect location regression to improve the accuracy of defect location. The experimental results show that the proposed algorithm can achieve a detection speed of 8.6 FPS on Raspberry Pi 4B, which can meet the needs of the textile industry.
Abstract: When the artificial potential field method is employed for unmanned aerial vehicle (UAV) path planning, there are often some problems, such as an unreachable target, a repeated motion trajectory, and a large path length. The traditional artificial potential field method fails to adjust the repulsion coefficient according to the specific information of the environment, while the existing improved methods cannot take into account the planning effect and planning time while adaptively adjusting the repulsion coefficient. To solve the above problems, this study proposes a UAV path planning method based on the adaptive repulsion coefficient with the help of deep learning. Firstly, the most suitable repulsion coefficient sample set in a specific environment is found by integrating a genetic algorithm and the artificial potential field method. Secondly, a residual neural network is trained by using the sample set. Finally, the repulsion coefficient adapted to the environment is calculated by the residual neural network, and then the artificial potential field method is used for path planning. Simulation experiments show that the proposed method solves the abovementioned problems in path planning with the artificial potential field method to a certain extent. It has excellent performance in planning effect and planning time and can well meet the requirements for current environment adaptation and rapid planning in UAV path planning.
Abstract: To address the problem that chemical oxygen demand (COD) is difficult to be measured on line during sewage treatment, this study proposes a soft sensing model based on a radial basis function (RBF) neural network. First, the process variables related to COD are selected as input variables by using the measured data of a sewage treatment plant. Second, the soft sensing model of COD in effluent is built on the basis of an RBF neural network. The center value, width value, and weight of the RBF neural network are optimized by an adaptive genetic algorithm improved sparrow search algorithm (AGAISSA). The accuracy of the soft sensing model is ensured by improving the sparrow position update formula and introducing the adaptive crossover and mutation operation in the genetic algorithm. Finally, the soft sensing model based on the RBF neural network is applied to the measured data of a sewage treatment plant for verification. The results show that the AGAISSA optimized RBF neural network model can accurately predict the COD in effluent and has high prediction accuracy.
Abstract: A water cleaning control system based on the reconfigurable program of statistical process control (SPC) is proposed to connect a water cleaning machine to an industrial network and realize intelligent automatic control. The system uses parameters and component characteristics of water cleaning equipment to construct the control protocol of the equipment and formulates an instruction judgment module according to the protocol instruction. In addition, it designs a process control module through SPC theory. The control protocol makes the cleaning program of the system capable of reconfiguration and networking. The judgment module ensures the security of reconstructed instructions. The process control module makes the cleaning process of the system able to dynamically adjust cleaned components. These functions enable the equipment to realize intelligent automatic control. Through the test, the average cleaning times of the system are reduced by 15%, and the utilization rate of water cleaning solutions is increased by about 5% compared with the original system. Furthermore, three functions have been expanded, which improves the utilization rate and intelligence level of the equipment, and the requirements of energy and water saving, multiple purposes, and networking are finally satisfied.
Abstract: Proximal policy optimization (PPO) is a stable deep reinforcement learning algorithm. The key process of the algorithm is to use clipped surrogate targets to limit step size updates. Experiments have found that when a clipping coefficient with optimal experience is employed, the upper bound of Kullback-Leibler (KL) divergence cannot be determined. This phenomenon is against the optimization theory of trust region. In this study, an improved PPO with double clipping boundaries (PPO-DC) algorithm is proposed. The algorithm adjusts the KL divergence based on two probability-based clipping boundaries and limits parameters to the trust region, so as to ensure that the sample data are fully utilized. In several continuous control tasks, the PPO-DC algorithm achieves better performance than other algorithms.
Abstract: Vehicle trajectory prediction can effectively reduce the collision risk caused by the sudden change of a vehicle trajectory, which is one of the key technologies to achieve safe driving. To address the problem that the traditional trajectory prediction algorithm lacks the analysis of the driver’s intention, this study proposes a vehicle trajectory prediction model that integrates generative adversarial networks and driving intention recognition. The model predicts vehicle trajectories under a generative adversarial network framework and introduces a deep neural network-based lane change intention recognition module to identify the driver’s lane change intention. A comparison test with LSTM, S-LSTM, CS-LSTM and S-GAN models is carried out on the public data set NGSIM. The experimental results show that compared with other trajectory prediction models, the CS-DNN-GAN model proposed in this paper has better prediction accuracy.
Abstract: Automatic text-speech alignment technology is widely used in speech recognition and synthesis, content production, and other fields. Automatic text-speech alignment aims to align speech with text in sentence, word, and phoneme units and obtain the time alignment information. Most of the recent alignment methods are based on automatic speech recognition (ASR). On the one hand, the alignment accuracy is limited by the word error rate (WER) of ASR. On the other hand, such methods cannot effectively align imperfect transcriptions. This study proposes a text-speech alignment method based on anchor and prosodic information. Through fragment annotation based on boundary anchor weighting, speech is divided into aligned and unaligned fragments. For unaligned fragments, this study extracts their prosodic information by a dual-threshold endpoint detection method and detects the boundaries of sentences. This approach reduces the dependence of ASR-based text-speech alignment on the speech recognition effect. Compared with the current advanced ASR-based text-speech alignment methods, the proposed method can improve alignment accuracy by more than 45% when the WER is 0.52 and by at least 3% when the degree of incomplete matching is 0.5.
Abstract: The weakly connected components of the time-evolving graph have been widely used in many areas, such as traffic network construction, information push of recommendation systems, etc. However, most methods for the weakly connected components ignore the impact of the non-uniform?memory?access (NUMA) architecture, that is, the high remote memory access delay leads to low execution efficiency. This study proposes a NUMA-based delayed sending method to find the weakly connected components of the time-evolving graph. It minimises the number of remote accesses and improves computational efficiency through reasonable data memory layout and controlling the number of exchanges between NUMA nodes. The experimental results show that the performance of the NUMA-based delayed sending method is better than the methods provided by the current popular graph processing systems Ligra and Polymer.
Abstract: Traditional chest-aided diagnosis systems have poor image feature extraction effects and low average accuracy in disease classification based on chest X-ray images. In view of these problems, a multi-level classification network that combines an attention mechanism and label correlation is proposed. The training of the network is divided into two stages. In stage one, in order to improve the feature extraction capability of the network, an attention mechanism is introduced, and a two-branch feature extraction network is constructed to realize the extraction of comprehensive features. In stage two, according to the correlation between labels and other issues in multi-label classification, a graph convolutional neural network is used to model the label correlation, which is then combined with the feature extraction results obtained in stage one, so as to achieve the multi-label classification task of diseases based on chest X-ray images. The experimental results show that the weighted average AUC of diseases by the proposed method on the ChestX-ray14 dataset reaches 0.827. Therefore, the method can assist doctors in diagnosing chest diseases and has certain clinical application value.
Abstract: In order to address the difference between visible light modality and thermal infrared modality and make full use of multimodal information to perform pedestrian detection, this study proposes a multimodal feature differential attention fusion pedestrian detection method based on YOLO. The method first uses the feature extraction backbone of the YOLOv3 deep neural network to extract multimodal features respectively. Second, the differential attention module of modal features is embedded between the corresponding multimodal feature layers to fully mine the difference information between modalities, and the difference feature representation is strengthened through the attention mechanism, so as to improve the quality of feature fusion. Then, the difference information is fed back to the multimodal feature extraction backbone to improve the network’s ability to learn and fuse multimodal complementary information. In addition, the multimodal features are fused in layers to obtain the multi-scale features. Finally, target detection is performed on the multi-scale feature layer to predict the probability and location of pedestrian targets. The experimental results on the public multimodal pedestrian detection datasets of KAIST and LLVIP show that the proposed multimodal pedestrian detection method can effectively address the difference between modalities and realize the full use of multimodal information. Furthermore, it has high detection accuracy and speed and is of practical application value.
Abstract: The power monitoring system is the most important production management system in the power industry. As an important part of the power monitoring system, the user station will become the main target of network attacks if it lacks grid binding. In order to perceive the network attack events on the subscriber station side in time, a method combining real-time detection and active defense of random domain names on the subscriber station side is proposed. A capsule network (CapsNet) combined with a long short-term memory (LSTM) network is used to classify the domain names extracted from the traffic data. When a random domain name is detected, instructions are sent to routers and switches to update their security policies or shut down the service interfaces of routers and switches to block network attacks through the remote terminal protocol (Telnet). The experimental results show that the use of the CapsNet combined with the LSTM classification algorithm can achieve an accuracy of 99.16% and a recall of 98% in random domain name detection. Through the Telnet, routers and switches can be linked to make active defense without interrupting services.
Abstract: In order to solve the identity authentication problem of network applications, the OAuth2.0 protocol has been widely used in the actual production environment. However, many systems use the OAuth2.0 standard unreasonably in their design, which results in many security flaws. This study analyzes the security problems of OAuth2.0 protocol in recent years, including the man-in-the-middle attack, authorization hijacking vulnerability, and CSRF vulnerability, and the study proposes a password-based Schnorr digital signature and OAuth2.0 strong identity authentication scheme for solving these security problems. Finally, the security of the scheme is analyzed. The results show that the scheme has excellent security and is easy to use.
Abstract: In the era featuring advanced technology and information explosion, how to accurately extract the required information from massive data has become the target studied by people. As one of the important ways to solve this problem, question-answering systems mainly retrieve and analyze existing data and information and finally return the answer to the question or other related information. In recent years, the revolutionary development of deep learning has brought considerable progress to question-answering systems. Sequence-to-sequence models, end-to-end models, and the recently popular pre-training have left unlimited development space for the question-answering systems, but these systems still face many challenges. This study first briefly introduces the development of the question-answering systems, then classifies these systems from three different perspectives, and expounds on the relevant data sets, evaluation indicators, and mainstream technologies of various question-answering systems. Finally, the paper discusses the problems faced by question-answering systems and their future development trends.
Abstract: By comprehensively considering the imbalance and sparsity of data in blended learning grade classification and prediction, this study proposes a blended learning grade classification and prediction model, namely, SMOTE-XGBoost-FM. Firstly, an equalization data set is sampled by SMOTE. In order to solve the problem of data sparsity, XGBoost is used to perform feature overlap on the sampled data, and then the leaf nodes of the generated tree are processed by one-hot encoding to generate high-order feature data. Finally, the data are input into a factorization machine (FM) for iterative training to obtain the optimal model. The experimental results show that the SMOTE-XGBoost-FM model achieves an accuracy of 92.7% in blended learning grade classification and prediction, which is 5.7% and 11.7% higher than that of single XGBoost and FM models, respectively. Therefore, it can effectively classify and predict students’ learning effects and provide a reference for improving teaching efficiency.
Abstract: Video salient object detection (VSOD) can continuously locate motion-related salient objects in video sequences by combining spatial and temporal information. Its core lies in how to efficiently describe the spatial and temporal features of moving objects. Existing VSOD algorithms mainly use optical flow, ConvLSTM, and 3D convolution to extract time domain features, but their continuous learning ability of temporal information is insufficient. Therefore, a robust spatial-temporal progressive learning network (STPLNet) is proposed to realize the efficient localization of salient objects in the video sequences. In the spatial domain, the method uses a U-shaped structure to encode and decode each video frame. In the temporal domain, it progressively encodes the features of the moving objects by learning the features of subject parts and deformation regions about the moving objects between frames in the video sequences. In this way, the method can capture the time correlation features and motion tendency of the objects. A series of comparative experiments are carried out on four public datasets, with 13 mainstream VSOD algorithms involved. The proposed model achieves optimal results on multiple indicators including maxF, S-measure (S), and MAE, and it has excellent real-time performance in running speed.
Abstract: Deep learning is currently a mainstream method for crack detection in pavement images, but it requires a large number of manually-annotated ground-truth images for training. However, in reality, it is time-consuming and laborious to obtain manually-annotated ground-truth images. This study proposes a method for crack detection in pavement images based on an improved generative adversarial network. The study regards crack detection in pavement images as a kind of anomaly detection problem based on image cross-domain transformation and uses a fixed-point generative adversarial network to automatically convert the crack image into a one-to-one corresponding crack-free image with supervision. Then, the study differentiates the original image and the generated image, and the salient objects in the difference image correspond to the crack detection results. The test results on the public dataset CrackIT show that the method in this study can achieve accurate crack detection without relying on the manually-annotated ground-truth images. In addition, the method achieves comparable performance to supervised deep learning methods in terms of precision, recall, and F1-score.
Abstract: Container virtualization is emerging in cloud computing due to its lightweight feature. Container live migration is the basis for many cloud management capabilities, which migrates a running container to another physical node with minimal downtime. Performance is the focus of container live migration research, but through a detailed analysis of existing container live migration systems, this study finds that there are still some problems affecting the performance, including low parallelism of dump, non-convergence of pre-copy policy, and low parallelism of root file system and running state migration. To solve these problems, this study proposes and designs three optimization strategies or algorithms including the resource awareness-based parallel dump mechanism, post-copy policy-based running state migration, and multi-priority-based parallel transfer scheduling algorithm. In addition, the paper realizes a high-performance container live migration system, namely, Dmigrate, based on Docker. Experimental results show that compared with the latest research, Dmigrate can effectively reduce downtime by 17.05%, and the total migration time is decreased by 24.33% on average.
Abstract: Medical images are helpful for the diagnosis, treatment, and evaluation of diseases. Accurate segmentation of organs in medical images is of great practical significance to assist doctors in diagnosis. Due to the low contrast between organ parts and surrounding tissues in medical images, the edges and shapes of different organs are very different, which increases the difficulty of segmentation. To solve these problems, this study proposes a semantic segmentation network for medical images based on a convolutional neural network and Transformer, which effectively improves the accuracy of semantic segmentation of medical images. The feature extraction part uses a ResNet-50 network structure, and a Transformer module is employed to expand the receptive field after feature extraction. In the process of up-sampling, multiple skip connection layers are added, and the feature extraction information of each stage is fully utilized to make the resolution close to that of input images. The experimental results on the segmentation dataset of gastrointestinal medical images prove that the proposed method can effectively segment organs and tissues in medical images and improve the segmentation accuracy.
Abstract: With the development of the smart Internet of Things (IoT) system, the type and number of applications in the IoT continue to increase. In mobile edge computing (MEC), mobile applications are accelerated by allowing mobile users to offload tasks to nearby MEC servers. This study simulates mobile edge scenarios by analyzing task attributes, user mobility, and delay constraints. In addition, according to the moving trajectory of users, the goal is modeled. Specifically, it aims to find an optimization model for MEC servers that satisfies the delay constraints and generates the minimum energy consumption during the offloading. The study also proposes a minimum energy offloading algorithm to find the optimal solution to this problem. The simulation results show that under the constraints, the proposed algorithm can find a MEC server that generates the minimum energy consumption in the moving trajectory of users, significantly reduce the energy consumption and delay during task offloading, and improve the application service quality.
Abstract: In the process of image acquisition, the image often contains certain noise information, which will destroy the texture structure of the image and thus interfere with semantic segmentation tasks. Most of the existing semantic segmentation methods based on noisy images adopt models featuring first denoising and then segmentation. However, they often lead to the loss of semantic information in denoising tasks, which thus affects segmentation tasks. To solve this problem, this study proposes a multi-scale and multi-stage feature fusion method for semantic segmentation of noisy images, which uses the high-level semantic information and low-level image information of each stage in the backbone network to enhance the semantic information of target contours. By constructing a staged collaborative segmentation denoising block, collaborative segmentation and denoising tasks are iterated, and then more accurate semantic features are captured. In addition, quantitative evaluation is carried out on PASCAL VOC 2012 and Cityscapes datasets. The experimental results show that the model still achieves positive segmentation results under the noise interference of different variances.
Abstract: To address the problems of missed detection of faces, the insufficient computing power of mobile platforms, and the limited hardware resources of face recognition applications under epidemic prevention and control, this study proposes an improved lightweight detection model for faces with masks based on YOLOv5. In this model, the C3 module in the original network is replaced with a lightweight C3Ghost module to compress the computations of the convolution process and the size of the model. Moreover, an attention mechanism is added to the backbone network to improve the feature extraction capability of the network, and the border regression loss function is improved to improve the speed and accuracy of detection. The experimental results indicate that the amount of calculation and parameters of the improved model are decreased by 29.79% and 33.33%, respectively, with the weight file size of only 2.8M. The improved model reduces the dependence on the hardware environment, and its detection rate reaches 96.6%. Compared with the existing models, it has outstanding advantages and can be effectively applied to face recognition.
Abstract: The factory environment is complex and changeable, with many dangerous areas, and illegal entry can bring serious harm to the life and health of workers. Considering the complex operation and poor recognition effect of traditional detection methods, this study proposes an alarm system for workers’ intrusion in dangerous areas on the basis of the improved YOLOv5s model. Firstly, the binocular ranging technology based on the SGBM algorithm is integrated into YOLOv5s object detection, and the trigger condition of spatial distance is added. Hence, the sound and light alarm will be triggered only when workers approach the camera within a certain range. Furthermore, the attention mechanism is introduced into YOLOv5s. Comparative experiments prove that the introduction of the CA module improves the average accuracy of mAP0.5 by 1.86%. The results show that this method can accurately identify the intrusion of a worker in dangerous areas and gives a sound and light alarm to remind the worker, which provides a new means for factory safety management.
Abstract: As clean energy, wind power plays an increasingly important role in improving China’s energy structure. Data on wind farm units and equipment may contain relevant privacy-sensitive information. Once the information is divulged, it will bring huge economic and legal risks to the wind farm. Federated learning (FL) is an important privacy-preserving computing technique, through which model training and inference are completed without transmitting raw data, so as to achieve joint computation among all participants without privacy disclosure and effectively deal with challenges in analyzing wind power data. However, significant communication overheads generated during FL computation have become a major performance bottleneck that has limited the application of the FL technique in wind power scenarios. Therefore, this study takes the typical FL algorithm, namely, XGBoost, as an example and deeply analyzes the communication problems in FL computation. In addition, the paper proposes a solution that RDMA shall be utilized as the underlying transport protocol and designs a set of high-performance FL platform communication libraries, which effectively improves the performance of the FL system.
Abstract: In a heterogeneous Hadoop cluster scenario, the hybrid use of erasure codes and replica storage modes, as well as the real-time computing capability difference of server nodes lead to the low efficiency of MapReduce job processing. To deal with this problem, this study implements a scheduling strategy that dynamically adjusts MapReduce job assignment in multi-concurrent scenarios according to data storage situations and the real-time load of nodes. This strategy dynamically controls the concurrent amount of tasks of each node by modifying data storage location strategies in the current Hadoop framework, so as to achieve more balanced resource allocation among jobs when multiple jobs are concurrent. The experimental results show that the scheduling mode proposed in this study can shorten the job completion time by about 17% and effectively avoid the starvation phenomenon faced by some jobs compared with the two default job scheduling strategies of Hadoop.
Abstract: As the core system of provincial meteorological business, the meteorological big data cloud platform (referred to as Tianqing) needs to work stably and efficiently for uninterrupted 7×24 hours. The Tianqing system has many operation modules and complex processing tasks. However, the traditional manual monitoring mode has low monitoring efficiency and cannot find faults or other problems existing in the business in time. In this study, the Java, Python, and Bash shell languages are used to develop a full-process monitoring and warning system for Tianqing business based on Enterprise WeChat. The system collects and formats the comprehensive status information generated during the business operation of each module of Tianqing into monitoring and warning information and finally sends it to the operation and maintenance personnel through the Enterprise WeChat, which realizes the quick perception of the operation status of each business operation module of Tianqing. The business operation effect of the system shows that the system is safe, reliable, and stable, and it can help operation and maintenance personnel to locate system faults in time and improve the efficiency of fault handling promptly. In addition, it has achieved positive application effects in Tianqing data monitoring and operation guarantee.
Abstract: In recent years, research on emotion recognition has no longer only focused on facial and voice recognition, and emotion recognition according to electroencephalogram (EEG)-based physiological signals is becoming increasingly popular. However, due to the incomplete extraction of feature information or the maladjustment of classification models, the classification effect of emotion recognition is poor. Therefore, this study proposes a hybrid model combining differential entropy (DE), convolutional neural network (CNN), and gated recurrent unit (GRU), namely, DE-CNN-GRU, to study EEG-based emotion recognition. The pre-processed EEG signals are divided into five frequency bands, and their DE features are extracted as preliminary features, which are then input to the CNN-GRU model for deep feature extraction and further classified by using Softmax. The hybrid model is tested on the SEED dataset. The result shows that the average accuracy obtained by the hybrid model is 5.57% and 13.82% higher than that obtained by using the CNN or GRU algorithm, respectively.
Abstract: In order to further improve the prediction accuracy of the air quality index, a hybrid genetic ant colony algorithm is proposed to optimize the back propagation (BP) neural network, so as to predict the air quality index. First, the pheromone distribution of the ant colony algorithm is initialized, and crossover and mutation operations of the genetic algorithm are performed if fitness conditions are not met. Then the state transition probability and pheromone concentration of the ant colony are calculated. When the fitness meets the conditions, the optimal results are used as the optimal weights and thresholds of the BP neural network to improve the shortcomings of a single BP neural network. Finally, historical daily data of the air quality index in Xi’an are utilized for verification, and the experiment shows that all evaluation indexes of the model proposed in this study have smaller errors than those of other comparative models and are more convincing in terms of prediction accuracy. Therefore, the proposed model can effectively predict the air quality index.
Abstract: The adversarial robustness of deep neural networks is of great significance in the field of image recognition. Relevant studies focus on the generation of adversarial samples and the robustness enhancement of defense models but lack comprehensive and objective evaluation. Thus, an effective benchmark to evaluate the adversarial robustness of image classification tasks is developed. The main functions of this system are list display, adversarial algorithm evaluation, and system optimization management. At the same time, computing resource scheduling and container scheduling are applied to ensure the evaluation task. This system can not only provide a dynamic import interface for a variety of attack and defense algorithms but also evaluate the advantages and disadvantages of the existing algorithms from all aspects in the confrontation between attack and defense algorithms.
Abstract: In the path planning of unmanned aerial vehicles (UAVs), the traditional algorithm has the disadvantages of complex computation and slow convergence, while particle swarm optimization (PSO) features simple principle, strong universality, and comprehensive search, which is mainly used in UAV route planning. As the conventional PSO algorithm is easy to fall into the local optimum, this study integrates the global extreme variation and acceleration terms based on the adaptive parameter optimization to balance the global and local search efficiency and avoid the population falling into “premature”. Through the test of a variety of benchmark functions, the results show that the improved PSO algorithm proposed in this study has faster convergence speed and higher convergence accuracy. In the example verification part, the flight scene features are first extracted, and the environment modeling is carried out based on the UAV performance constraints. Then multiple constraints and the expected minimum flight time are converted into penalty functions. With the minimization of penalty functions as the objective, the route planning model is constructed, and the improved PSO algorithm is adopted to solve the problem. Finally, the effectiveness and practicability of the improved PSO algorithm are verified by comparative simulation.
Abstract: Flame image recognition faces a large model size, low accuracy, and poor real-time performance in edge equipment and mobile terminal equipment environments. In order to solve these problems, firstly, ShuffleNetV2 is selected as a lightweight backbone neural network to ensure the real-time performance of the model. Secondly, a new space and channel dual attention module (SCDAM) is designed to analyze the relevance of channels and spaces, and different weights are given according to the importance of different features, so as to effectively improve the model accuracy. Then, in order to enrich the extracted features on the spatial scale, a multi-scale feature fusion module is designed to enhance the adaptability of the network to different scales. Finally, the SCDAM and multi-scale module are introduced into ShuffleNetV2, and the model parameters are optimized by transfer learning, so as to further improve the model accuracy. With only a slight increase in the amount of parameters and calculations, the accuracy of the proposed algorithm is 3.2% higher than that of ShuffleNetV2, and the single inference takes only 8.7 ms. Experiments show that the proposed algorithm is more suitable for conditions with limited computing resources, such as the identification and monitoring of gunpowder flame.
Abstract: As small targets in the early stage of a fire are difficult to detect during flames and smoke detection of fires, this study proposes an improved YOLOX-nano (ASe-YOLOX-nano) object detection algorithm based on natural exponential loss (eCIoU). Firstly, a new object detection function, the eIoU loss function, is proposed to replace the traditional IoU loss, which solves the problems of no intersection between the prediction box and the real frame in small target detection and the inability to react to the influence of width and height. Secondly, the attention module is introduced in the network model to vaguely locate the target position in the early stage of the network and improve the accuracy of the detection of targets, especially small targets, in the later stage of the network. In addition, the soft pooled spatial pyramid pooling structure is employed to extract spatial feature information of different sizes, which can improve the robustness of the model for spatial layout and object degeneration. In this way, sufficient features can be extracted when the target is small. Moreover, the Mosaci enhancement technology is used to preprocess the dataset to improve the generalization ability of the model for further improvement in network performance. The comparative verification of the target data set shows that the mAP index reaches 70.07%, which is 3.46% higher than that of the original model, and the model enjoys accuracy of flame and smoke detection of 84.66% and 74.56%, respectively, and a stable FPS of 73, which has better fire detection ability than the traditional YOLOX-nano algorithm.
Abstract: Traditional three-dimensional (3D) dense captioning methods have problems such as insufficient consideration of point-cloud context information, loss of feature information, and thin hidden state information. Therefore, a multi-level context voting network is proposed. It uses the self-attention mechanism to capture the context information of point clouds in the voting process and utilizes it at multiple levels to improve the accuracy of object detection. Meanwhile, the temporal fusion of hidden state and attention module is designed to fuse the hidden state of the current moment with the attention result of the previous moment to enrich the information of the hidden state and thus improve the expressiveness of the model. In addition, a “two-stage” training method is adopted in the model, which can effectively filter out the generated low-quality object proposals and enhance the description effect. Extensive experiments on official datasets ScanNet and ScanRefer show that this method achieves more competitive results compared to baseline methods.
Abstract: Cognitive impairment in the elderly is becoming a major threat to their life quality, but preventive measures, treatment technologies, and health care models are still immature for the elderly with cognitive impairment. Additionally, there is a lack of data systems for cognitive impairment in the elderly that can store these medical data in a complete and disaggregated manner. These lead to inaccurate diagnosis of cognitive impairment, delayed treatment of cognitive impairment, and unavailable appropriate medical care for cognitive impairment patients. To address these problems, this study designs a B/S-based multidimensional data management system for cognitive impairment in the elderly. The system takes advantage of the FastDFS distributed file storage system to ensure the security and stability of the system data. The recursive tree structure is adopted to extract the tabular data and speed up the screening process. The system is compatible with current major browsers.
Abstract: In water surface garbage detection, large differences occur in target shape and scale, and it is difficult to distinguish the background and the small target. Thus, this study proposes the SPMYOLOv3 detection algorithm to identify surface garbage. Firstly, massive surface garbage datasets are collected and annotated, and an improved k-means clustering method is applied to generate the priori boxes that better match the datasets. Secondly, the SE-PPM module is added after the backbone network of YOLOv3 for strengthening the feature information of the target, ensuring that the target scale remains unchanged and the global information is preserved. The multidirectional FPN is then applied to fuse the feature maps of different scales so that the feature maps after fusion contain richer context information. Finally, the Focal Loss is adopted to compute the confidence loss of negative samples, which alleviates the imbalance of positive and negative samples in YOLOv3. The modified algorithm is tested on the water surface garbage dataset, and the results show that the accuracy of the modified algorithm is 3.96% higher than that of the original YOLOv3 algorithm.
Abstract: The recommendation algorithm based on the graph neural network generates the feature representation of nodes by obtaining knowledge from graphs, which improves the interpretability of recommendation results. However, as the original data scale of the recommendation system continually expands, a large amount of text data containing semantic information has not been effectively used. Additionally, the graph neural network does not distinguish key nodes when fusing the neighbor information in the graph, making it difficult for the model to learn high-quality entity features, which in turn leads to a decrease in the quality of recommendation. This study combines the graph neural network with a semantic model and proposes a recommendation algorithm based on the graph neural network which integrates semantic information and attention. This algorithm processes entity-related text information based on the SpanBERT semantic model and generates feature embeddings containing semantic information. It also introduces the attention mechanism into the process of influence propagation and fusion based on user social relations and user-item interactions to effectively update user and item entity features. The comparative experimental results on public datasets show that the proposed method is better than the existing benchmark methods in all indicators.
Abstract: Compared with traditional logistics warehouses, many automated warehouses use automatic guided vehicles (AGVs) instead of workers to sort the goods, which changes the working mode from “man to goods” into “goods to man”. This change not only liberates the labor of workers but also combines the mechanization and automation of automated warehouses, greatly improving working efficiency. Path planning is an important part of AGVs in the process of sorting goods in automatic warehouses. For the path planning of AGVs, the traditional A* algorithm is improved because the paths planned by the traditional A* algorithm are too long and not smooth enough and have large turning angles. In view of the above defects, the method of dynamic weighting and changing the search neighborhood is proposed to improve the traditional A* algorithm, which reduces the search nodes and raises the search speed. At the same time, the path planned by the improved A* algorithm is smoothed by higher-order Bessel curves many times, which lowers the number of turning points. Finally, the comparison of three groups of simulation experiments proves that the improvement proposed in this study is of the reference value.
Abstract: Generating short and readable regular expressions from finite automata is an important topic in computer theory. In the classical regular expression generation algorithms, the state sequence is the key factor that affects the quality of regular expressions. To search for excellent state sequences quickly and efficiently, this study takes the theory of the carnivorous plant algorithm as the core, combines the ideas of other heuristic algorithms for design and optimization, and proposes a state sequence search method based on the carnivorous plant algorithm. Through experiments, this method is compared with some existing search algorithms using heuristic rules. The experimental results demonstrate that the proposed state sequence search method is superior to other algorithms, and the length of the generated regular expressions is significantly shorter than that of other heuristic algorithms. For example, compared with the results of the DM algorithm, the length can be shortened by more than 20% with the increase in the order of automata, and compared with the results of the random sequence algorithm, the length can be shortened by several orders of magnitude.
Abstract: In recent years, research on the named entity recognition of poetry in digital humanities is emerging, but few studies have been conducted with regard to the feature expressiveness of character features, word segmentation accuracy, and the effectiveness of domain-specific knowledge in poetry texts. According to the characteristics of Chinese pictographs and the particularity of poetry texts, a recognition method of named poetry entities with a feature enhancement unit and a feature extraction unit is proposed, which integrates multiple features such as characters, radicals, sounds, and metrical rules. The method presents the knowledge vectors obtained from the knowledge triples of tune pattern titles through the ANALOGY model as the knowledge vectors of tune pattern titles. Then, the radical vector, character vector, metrical rule vector, sound vector, and knowledge vector of tune pattern titles are deeply fused through the bidirectional long short-term memory network and attention mechanism models. In this way, the recognition method of named poetry entities fusing multi-features is constructed. The results of comparative experiments and ablation experiments on the self-made corpus of Translation of Among Flowers (Hua Jian Ji) (《花间集全译》) show that the proposed method can effectively use multi-features to improve the recognition performance of named entities, and its F1 score reaches 85.63%, which means it completes the recognition task of named poetry entities.
Abstract: As one of the most challenging problems in combinatorial optimization, the traveling salesman problem has attracted extensive attention from the academic community since its birth, and a large number of methods have been proposed to solve it. The ant colony algorithm is a heuristic bionic evolutionary algorithm for solving complex combinatorial optimization problems, which is effective in solving the traveling salesman problem. This study introduces several representative ant colony algorithms and makes a literature review of the improvement, fusion, and application progress of ant colony algorithms to evaluate the development and research achievements of different versions of ant colony algorithms in solving the traveling salesman problem in recent years. Moreover, the improved ant colony algorithms are summarized in categories in terms of the framework structure, setting and optimization of algorithm parameters, pheromone optimization, and hybrid algorithms. The research provides an outlook and basis for the application of ant colony algorithms to solve the traveling salesman problem and further develop the research content and focuses of other fields.
Abstract: Entity alignment aims to find equivalent entities located in different knowledge graphs and is an important step for knowledge fusion. Currently, mainstream entity alignment methods are those based on graph neural networks. However, they often rely too much on the structural information of graphs, as a result of which models trained on specific graph structures cannot be applied to other graph structures. Meanwhile, most methods fail to fully utilize auxiliary information, such as attribute information. In response, this paper proposes an entity alignment method based on a graph attention network and attribute embedding. The method uses the graph attention network to encode different knowledge graphs, introduces an attention mechanism from entity application to attribute, and combines structure embedding and attribute embedding in the alignment stage to improve the effect of entity alignment. The proposed model is verified on three real-world datasets, and the experimental results show that the proposed method outperforms the benchmark methods for entity alignment by a large margin.
Abstract: The detection and recognition of ship numbers are of great significance for the intelligent management of ports and can solve the time-consuming and labor-intensive problems caused by the traditional manual supervision of fishing boats. Since the ship number plates feature non-uniform hanging positions, background colors, and numbers of characters, this study proposes a two-stage detection and recognition method with two models. First, the study introduces a DP-DBNet ship number location detection model that combines dual path networks (DPN) with a differentiable binarization network (DBNet). Secondly, the study presents an MHA-CRNN ship number recognition model that combines the multi-head-attention mechanism (MHA) with the improved convolutional recurrent neural network (CRNN). Finally, this study uses the data from the new modern smart fishing port project in Zhifu District of Yantai and carries out an algorithm comparison experiment analysis. The experimental results show that the two-stage recognition method with two models can make the recognition accuracy rate of the ship number reach 76.39%, which fully proves the effectiveness and application value of the model in marine port management.
Abstract: Manual planning of city marathon routes has low efficiency. In view of this, this study adopts a greedy and backtracking algorithm to carry out intelligent planning of a city marathon route. The specific method is described as follows. A road network connected by the topological relationship of longitude and latitude coordinate points is built through the urban road network information, and a traversal search is performed by the greedy and backtracking algorithm on the coordinate points. In addition, according to the special requirements of the city marathon route, strategies are adopted, such as direct approximation, heuristic distance, heuristic approach, and direction estimation, so as to realize the intelligent planning of the route. On this basis, a marathon route evaluation method is proposed, which integrates five dimensions including POI heat value, road width suitability, route smoothness index, comfort for turning, and POI density. Finally, a comparative analysis of artificial and intelligent route planning for Beijing and Hefei marathons is carried out. The results show that the proposed method can realize fast and efficient marathon route planning.
Abstract: In order to realize the automatic optimization of the optimal parameters of grayscale image enhancement, an adaptive image enhancement method based on an improved squirrel search algorithm is proposed. A bilateral search strategy is introduced into the position updating of the squirrels on normal trees to increase the likelihood of obtaining an optimal solution. A cyclone foraging strategy is used to update the position of the squirrels on acorn trees to improve the convergence rate and search accuracy of the algorithm. In addition, the proposed squirrel search algorithm with bilateral search and cyclone foraging (BCSSA) is compared with the bat algorithm (BA), whale optimization algorithm (WOA), basic squirrel search algorithm (SSA), and two improved SSAs on CEC 2017 test suite. The results indicate that BCSSA has higher stability and faster convergence rate. Finally, the proposed BCSSA is applied to grayscale image enhancement, and its performance is compared with that of the classical histogram equalization method and SSA in terms of four evaluation indicators, which thus validates the superiority of BCSSA.
Abstract: It is essential for banks to accurately predict whether clients will use their credit and analyze key factors influencing credit utilization after these clients have been approved for credit, so as to improve their client service level and profitability. Currently, machine learning algorithms are rarely applied to credit utilization prediction, and there is a lack of research on model interpretability in the financial credit utilization field. Therefore, this study proposes an interpretative TreeSHAP credit utilization prediction model based on CatBoost. Specifically, a credit utilization prediction model is constructed by CatBoost and is compared and optimized by using three hyperparameter optimization algorithms. Then, the model is experimentally compared with baseline models in terms of four main performance metrics. The results show that the model optimized by the TPE algorithm outperforms other models. Finally, the interpretability of the model is enhanced locally and globally by the TreeSHAP method. Furthermore, factors influencing client credit utilization are interpretively analyzed, so as to provide a decision-making basis for banks to make accurate marketing to clients.
Abstract: To identify the running fault of motors quickly and effectively from the temperature data collected by thermal imagers, this paper combines dropout, nonlinear wavelet transform coefficient enhancement (NLWTCE), and convolutional neural network (CNN) algorithm to identify the motor image. Firstly, the image dataset of the motor is established according to the data collected by the thermal imager and the data image is enhanced by nonlinear wavelet transform (NLWT). Then an improved CNN (ICNN) model is built to identify the image with the extracted features as the final recognition features. Finally, compared with the normal motor images, the faulty motor images are effectively and accurately identified. The experimental results show that the ICNN model not only has a high recognition accuracy but also further simplifies the complex extraction of image features. The validity and reasonableness of the method are verified, and the method is suitable for engineering application.
Abstract: Automatic dependent surveillance-broadcast (ADS-B) is an important part of the new generation air traffic management system of civil aviation. As the protocol does not have data encryption and authentication, it is vulnerable to data attacks. To accurately detect ADS-B data attacks, based on the timing of ADS-B data, this study proposes a convolutional neural networks-long short-term memory (CNN-LSTM) anomaly detection model based on attention mechanism. Firstly, CNN is adopted to extract the features of ADS-B data, and then the feature vectors are input into the LSTM in the form of time series. Finally, the attention mechanism is applied to optimize the network parameters to realize the prediction of ADS-B data, and the anomaly detection is carried out by calculating the prediction error. Experiments show that the model can well detect four types of abnormal data and has higher accuracy and F1 score than other machine learning methods.
Abstract: Manufactured sand refers to artificial sand whose particle size is less than 2.36 mm after the repeated crushing of gravels by sand-making machines. In experiments, stone powder and mud contents in the manufactured sand are called fine powder content, which represents the cleanliness of the manufactured sand. In this study, a method for predicting the fine powder content in the manufactured sand based on the XGBoost network is proposed. First, a completely closed image acquisition device is used to collect images of a solution made of fine powders in the manufactured sand, so as to guarantee that the outside light will not affect shooting. Then pre-treatment is carried out, such as picture cropping, RGB value reading, and LCH color space shifting, and an XGBoost network model is built. Through the Bayes principle, loop iteration of parameters is conducted, and the model is optimized, so as to make the r2_score of the model higher and finally predict the fine powder content in the manufactured sand. The results show that the r2_score of the data predicted by this model can reach 0.967 762. In addition, the r2_score predicted by the traditional multiple linear regression models, BP neural network, and traditional XGBoost network is 0.896 144, 0.914 598, and 0.950 670. In contrast, the prediction accuracy of the proposed model is significantly improved. In practical application, this method can shorten the measurement time and simplify the measurement steps of the fine powder content in the manufactured sand. Therefore, it is a new method for predicting the fine powder content in manufactured sand.
Abstract: In order to correct the nonstandard riding postures of riders, a parametric modeling method for realizing standardized riding is proposed. Firstly, a manikin and a bicycle model are created, and parameters of the bottom layer, middle layer, and high layer are defined, so as to realize model parameterization. Secondly, force analysis during the riding is carried out, and a kinetic model is established, so as to ensure that the virtual riding conforms to the natural motion law. Finally, the constraint relationship between human upper and lower limb parameters and bicycle parameters is established to realize the coordinated movement of human joints, and the motion during the riding is simulated. The simulation results show that this method can provide correct posture guidance for riders.
Abstract: Non-intrusive load monitoring (NILM) is an important part of intelligent power utilization and energy saving techniques and has attracted extensive attention. Due to the superior performance of newly-developed deep learning methods in various tasks in recent years, some representative deep learning methods have been successfully applied to the load decomposition task in NILM. To systematically summarize the research status and progress of deep learning methods applied to NILM, this study focuses on analyzing and summarizing the research literature on deep learning based NILM in recent years. Firstly, the NILM framework is outlined, and then the feature extraction method and the public data set of NILM are introduced. In addition, the load decomposition methods based on deep learning in NILM are analyzed and summarized. Finally, the study points out several challenges in this field and provides an outlook on its opportunities and future research directions.
Abstract: As the core technology to realize the information and digital transformation of the construction industry, BIM (building information modeling) technology has high research value in the whole life cycle of railway construction. In the design of railway communication machinery rooms, station yards, and sections, the spatial morphology of railway communication entity, such as spatial position, shape, size, and relationship, is described digitally. According to railway communication design specifications, relevant railway BIM standards, and professional actual design requirements, the digital engineering design system for railway communication is studied and developed. Supported by spatial morphological data and based on the decomposition standard of railway engineering entity structure, the system realizes the intelligent layout of indoor cabinet equipment of railway communication machinery, the path planning of station communication trench cables, and the accurate layout of section communication information points in a three-dimensional environment. Based on the digital engineering model and the basic principle of graph theory, the system obtains the logical relationship from the digital engineering model and generates the communication logical network diagram. Verified by the actual project, the system has greatly improved the design efficiency and accuracy of railway communication digital engineering, realized the delivery and application of digital achievements of railway commu-nication engineering from the source of the project, and promoted the upgrading of technologies and innovation of digital modes in the whole process of railway communication engineering project.
Abstract: Negotiation refers to the process in which people communicate with each other on certain topics to reach an agreement. Automated negotiation aims to reduce negotiation costs, improve negotiation efficiency, and optimize negotiation results by using negotiating agents. In recent years, deep reinforcement learning techniques have been applied to the field of automated negotiation with good results. However, there are still problems such as the long training time of agents, dependence on specific negotiation domains, and insufficient utilization of negotiation information. Therefore, this study proposes a negotiation strategy based on the TD3 deep reinforcement learning algorithm, which reduces the exploration cost of the training process through pre-training and improves the robustness of the negotiation strategy by optimizing the state and action definitions, so as to adapt to different negotiation scenarios. In addition, it makes full use of the interaction information of the negotiation by multi-head semantic neural network and opponent preference prediction module. The experimental results show that the strategy can perform the negotiation task well in different negotiation environments.
Abstract: A virtual sports interaction system based on real-time video perception is proposed in response to the problems that traditional sports are limited by venues and equipment in the context of ongoing COVID-19 response, and the related products in the market are expensive and not scalable. The system is designed with a video data acquisition module and a human joint point extraction module, which can acquire human joint point coordinates in combination with OpenPose and capture human gestures and body movements in real time. The action semantic understanding module includes motion action understanding and drawing action understanding. The former recognizes the motion action semantics depending on the relative position relationship of the limb joints in motion. The latter generates the drawing action trajectories of wrist joints as sketch images, uses AlexNet to recognize and classify them, and resolves them into the corresponding drawing action semantics. The classification accuracy of the model is 98.83% in edge-side devices. A Unity-based sketch game application is used as the visual interaction interface to realize motion interaction in a virtual scene. The system adopts the interaction mode of real-time video perception to achieve home exercise and fitness without other external devices, which is more participatory and interesting.
Abstract: The stock market is an important part of the financial market, and it is of great importance for stock price prediction. Meanwhile, deep learning has powerful data processing capability to solve the problems caused by the complexity of financial time series. In this regard, this study proposes a hybrid neural network model (ATLG) that combines a self-attention mechanism, a long short-term memory (LSTM) network, and a gated recurrent unit (GRU) for stock price prediction. The experimental results show the followings: (1) The ATLG model has higher accuracy than LSTM, GRU, RNN-LSTM, and RNN-GRU models. (2) The introduction of the self-attention mechanism makes the model more focused on the information of stock characteristics at important time points. (3) Comparison reveals that the two-layer neural network plays a more distinct role. (4) The backtesting with the moving average convergence and divergence (MACD) indicator achieves a 53% return, which is higher than the return of CSI 300 in the same period. The results prove the effectiveness and practicality of the model in stock price prediction.
Abstract: Accurate recognition of crop pests is essential for timely crop protection and treatment. However, crop pests in natural environments are small in size and have almost the same color as the environments. Moreover, crop pest images are affected by various factors such as noise and complex backgrounds. Therefore, it is difficult for existing crop pest recognition models related to deep learning to balance the requirements of recognition accuracy and robustness and be deployed on mobile devices with limited computational resources and low performance. In this study, ShuffleNetV2 0.5×, which has the fewest model parameters in the ShuffleNetV2 network structure, is selected as the benchmark network, and a lightweight crop pest recognition model based on high-order residual and attention mechanism (HOR-Shuffle-CANet) is designed. Specifically, the high-order residual in the early stage can provide rich pest features for the subsequent network layer, which significantly improves the recognition accuracy of the model. The coordinate attention (CA) mechanism can further suppress the background noise and focus on the key information about crop pests, which effectively enhances the robustness of the model. The bi-tempered logistic loss function with label smoothing regularization (LSR) can solve two shortcomings of logistic loss functions in training noisy data sets and make the model more adaptable to noise. The experimental results show that the HOR-Shuffle-CANet model achieves a recognition accuracy of 91.22% on the test dataset of ten types of common crop pest images in natural scenes, which is 3.54 percentage points higher than the benchmark network. On the basis of maintaining lightweight computing, its recognition accuracy is also higher than that of the existing classical convolutional neural networks such as AlexNet, VGG-16, GoogLeNet, Xception, and ResNet-34, as well as lightweight network models such as MobileNetV3-Small, EfficientNet-B0, etc. Due to its high recognition accuracy, strong robustness, and excellent anti-interference performance, the proposed model can meet the practical application requirements of crop pest recognition.
Abstract: Automatic program repair is an effective technology for ensuring software quality and improving development efficiency. At present, most automatic repair tools use test cases as the final method of patch correctness verification. However, program can barely be fully tested by limited test cases. Consequently, patch sets generated by automatic repair tools contain a large number of incorrect patches. To identify such patches, this study identifies the effectiveness of repair patches by comparing the execution paths of successful tests before and after defect repair and the methods of test case generation to solve the low accuracy problem of automatic repair tools. When the proposed method is applied to evaluate 132 patches generated by six classic repair tools, it successfully excludes 80 incorrect patches, without excluding correct ones. This result shows that the proposed method can effectively exclude incorrect patches and improve the accuracy of automatic repair tools.
Abstract: Exploring and protecting fish is an important part of maintaining the balance of the marine ecological environment. However, the complex underwater environment affected by light, water quality, and occlusions makes it difficult to identify blurred fish images captured underwater and consequently restricts the speed and accuracy of underwater fish target detection. To solve the above problem, this study proposes a marine fish identification model based on improved fully convolutional one-stage object detection (FCOS). Specifically, the model takes the one-stage FCOS algorithm as the basic structure and uses the lightweight MobileNetv2 as the backbone network, which not only ensures the detection accuracy but also improves the detection; then, an adaptive spatial feature fusion (ASFF) module is introduced to avoid the inconsistency in scale features and improve detection accuracy; finally, the center-ness branch is introduced into the regression branch, and the generalized intersection over union (GIoU) loss is introduced to improve detection performance. Regarding the experimental dataset, the pictures in the public dataset Fish4Knowledge (F4K) and video frame screenshots are utilized, and the model with the optimal training performance is selected for evaluation. The results show that the average detection accuracy of the proposed new model on the above datasets is 99.79% and 99.88%, respectively. Compared with the original model and other detection models, the proposed model provides higher detection accuracy and identification speed. The model in this paper can provide a reference for marine fish identification.
Abstract: Taking the point cloud data from unmanned aerial vehicle (UAV) images of expressways as the research object, this study proposes a road green belt segmentation algorithm based on double decision factors. For this purpose, the point cloud data is serially down-sampled to retain as many point cloud feature points as possible in addition to reducing the number of point clouds; then, orthorectification of the down-sampled point cloud data is performed; finally, a point cloud segmentation algorithm featuring double decision with the normal vector angle and random sample consensus (RANSAC) plane segmentation is proposed, and accurate segmentation of the green belts in expressways is thereby achieved. The information on the environment of expressways is ultimately segmented with the green belt boundary extraction algorithm. Taking the point cloud from the UAV images of the Fengxiang section of G85 Expressway as the experimental data, this study verifies the proposed algorithm, the segmentation algorithm based on the normal vector angle, and the one based on RANSAC plane fitting. The experimental results show that the road green belt segmentation algorithm based on double decision factors can better resist the interference from environmental noise and outliers, effectively filter the high curvature points on the road surface, and ultimately obtain better extraction results.
Abstract: With the rapid growth of internet finance and electronic payment business, resulting personal credit problems are also increasing. Personal credit prediction is essentially an imbalanced binary sequence classification issue. Such an issue is faced with a large size and high dimension of data samples and extremely imbalanced data distribution. To effectively distinguish the credit situation of applicants, this study proposes a personal credit prediction method based on feature optimization and ensemble learning (PL-SmoteBoost). This method involves the construction of a personal credit prediction model within the boosting ensemble framework. Specifically, data initialization analysis with the Pearson correlation coefficient is conducted to eliminate redundant data; some features are selected with the least absolute shrinkage and selection operator (Lasso) to reduce data dimension and thereby lower high dimensional risks; linear interpolation among the minority classes in the dimension-reduced data is carried out by SMOTE oversampling to solve the class imbalance problem; finally, to verify the effectiveness of the proposed algorithm, this study takes the algorithms commonly used to deal with binary classification issues as comparison methods and tests the algorithms with the high dimensional imbalance datasets downloaded from the open databases of Kaggle and Microsoft. With the area under the curve (AUC) as the algorithm evaluation index, the test results are analyzed by the statistical test method. The results show that the proposed PL-SmoteBoost algorithm has significant advantages over other algorithms.
Abstract: This study is conducted to study the influence of computer virus spread on the security situation of network systems. It analyzes the relationship between the SIR epidemic spread model and computer network security and proposes a SIPM model for network security situation prediction. Specifically, the SIPM model adds the memory function of nodes for different virus propagation, supports the independent propagation of multiple viruses in the network at the same time, and improves the dynamic propagation equation on the basis of the SIR model. It allows the independent setting of the infection ability of viruses to different device nodes and that of the resistance of device nodes to different viruses, which is closer to the real network environment. The experimental analysis uses a typical campus network architecture for simulation, and the results show that the model can analyze and predict the network security situation from many aspects.
Abstract: To solve the problems, such as low prediction accuracy, in tasks of aeroengine remaining useful life (RUL) prediction due to insufficient representative feature extraction, this study proposes an RUL prediction method based on multi-feature fusion for aeroengines. Exponential smoothing (ES) is performed to reduce the interference noise in the original data and thereby obtain relatively stable feature data. The time series features of the feature data are extracted by the bidirectional long short-term memory (Bi-LSTM) network and then assigned weights through the multi-head attention mechanism (Multi-Attention). A convolutional long short-term memory (Conv-LSTM) network is designed to extract the spatio-temporal features of the feature data. Then, the handcrafted features of the feature data are extracted, and weights are calculated from the Softmax functions. A feature fusion framework is designed to fuse the above features, and RUL prediction is finally achieved by fully connected network regression. The commercial modular aero-propulsion system simulation (C-MAPSS) dataset is used to simulate and verify the proposed model. Compared with Bi-LSTM and other models, the proposed model achieves higher prediction accuracy and better adaptability.
Abstract: PM2.5 is an important indicator for measuring the concentration of air pollutants, and monitoring and predicting its concentration can effectively protect the atmospheric environment and further reduce the harm caused by air pollution. As automatic air quality monitoring stations are constructed on a large scale, the air quality prediction model built by traditional machine learning can no longer meet the current needs. This study proposes a Gaussian-attention prediction model based on the multi-head attention mechanism and Gaussian probability estimation and utilizes the data from a monitoring station in Shenyang for training and tests. Because PM2.5 concentration is affected by other air quality data, this model uses the information alignment of hierarchical time stamps (week, day, and hour) of air quality data as input and extracts the time-series correlation features of different subspaces with the multi-head attention mechanism. More complete and effective feature information is thereby obtained, and prediction results are then acquired by Gaussian likelihood estimation. A comparison with multiple benchmark models is conducted, and the mean squared error (MSE) and mean absolute error (MAE) of the proposed Gaussian-attention prediction model are respectively 21% and 15% lower than that of the DeepAR model. Effectively improving prediction accuracy, the proposed model can accurately predict PM2.5 concentration.
Abstract: The security of electric energy plays an important role in national security. With the development of power 5G communication, a large number of power terminals have positioning demand. The traditional global positioning system (GPS) is vulnerable to spoofing. How to improve the security of GPS effectively has become an urgent problem. This study proposes a GPS spoofing detection algorithm with base station assistance in power 5G terminals. It uses the base station positioning with high security to verify the GPS positioning that may be spoofed and introduces the consistency factor (CF) to measure the consistency between GPS positioning and base station positioning. If CF is greater than a threshold, the GPS positioning is classified as spoofed. Otherwise, it is judged as normal. The experimental results show that the accuracy of the algorithm is 99.98%, higher than that of traditional classification algorithms based on machine learning. In addition, our scheme is also faster than those algorithms.
You are the firstVisitors
Copyright: Institute of Software, Chinese Academy of Sciences Beijing ICP No. 05046678-3
Address：4# South Fourth Street, Zhongguancun,Haidian, Beijing,Postal Code：100190
Phone：010-62661041 Fax： Email：csa (a) iscas.ac.cn
Technical Support：Beijing Qinyun Technology Development Co., Ltd.